Part 1 of a five-part Project Milk Carton Analysis
PROLOGUE
SEALED AT MIDNIGHT
At 11:59 p.m. Eastern Time on Monday, July 13, 2026, a window closed in Washington. It closed quietly, the way federal windows do — no press conference, no floor vote, no hearing.
It was the application deadline for funding opportunity HHS-2026-ACF-ACYF-CA-0037, a $6 million program run by the Administration for Children and Families inside the Department of Health and Human Services.
The applications that arrived before midnight are sealed inside the federal grants system now. The public does not get to see who applied.
What the public does get to see — because it is printed in the funding notice itself — is what the money is for.
The notice anticipates up to ten awards of $400,000 to $600,000 each, open to state, territorial, and tribal child-welfare agencies, to pilot predictive analytics in child welfare. In the agency’s own language, the program contemplates tools to assist decisions during hotline intake, systems that flag high-risk cases during investigations, models that generate risk scores or forecast maltreatment, and analytics used to identify, target, and retain foster and adoptive placements.
Awards are expected around September 30, 2026. Two more application cycles follow, in June 2027 and June 2028.
Read cold, that is a technology grant.
Read against the ten-year public record — which is what this analysis does — it is something else: a federal offer to support statewide predictive-analytics projects in a field whose documented history includes a tool in Illinois that assigned more than 4,100 children a 90-percent-or-greater probability of death or serious injury while failing to flag highly publicized cases in which children later died; a Los Angeles pilot that generated 3,829 false positives; an Oregon tool terminated weeks after an Associated Press investigation into the Allegheny model on which Oregon’s tool was based; a New York City model with 279 variables that families, their lawyers, and caseworkers are not told has flagged a case for additional review; and a Pittsburgh-area tool that became the subject of disability-discrimination complaints filed with the Justice Department in 2022.
Associated Press reporting in January 2023 said Justice Department civil-rights attorneys were scrutinizing the Allegheny tool. We found no public disposition of those complaints as of publication.
The machines score families.
Oversight of the machines, however, is inconsistent. Evaluations exist, but there is no uniform requirement for independent, recurring audits across jurisdictions.
And as of midnight on July 13, the applications to build more of them are in.
This is the sixth module in Project Milk Carton’s series on the machinery of American child welfare. Like its predecessors, it runs on one rule, stated up front: every factual claim, number, name, case, document, and quotation in this article is drawn from publicly available sources — federal statutes and regulations, the funding notice itself, peer-reviewed research, federal court records, government reports, identified investigative reporting, and nonprofit financial filings, including filings held in Project Milk Carton’s own public-data archive.
Every figure is attributed.
Where the analysis is ours, we say so. Where we could not verify something, we tell you that too. Where we draw an inference from the documented record, we identify it as analysis.
Project Milk Carton produces transparency research from public data.
Because the record is too large for one responsible article, this one is being published in five parts.
Part I begins with the machine itself: what these systems actually predict, what data they consume, and why those inputs matter.
Part II follows the record through Illinois, Los Angeles, Oregon, and Allegheny County, then examines the researchers and evidence base behind the field.
Part III follows the vendor and the federal money.
Part IV examines disability, civil-rights complaints, and the new federal funding notice line by line.
Part V asks the accountability question and lays out four reforms.
Then we will come back to the envelope.
Because the story does not end at midnight.
It starts there.
THE MACHINE
WHAT THE SCORE ACTUALLY PREDICTS
The substitution at the heart of the field
Begin with the single most important fact about child-welfare risk scoring, the one the brochures rarely lead with: the flagship models do not directly measure whether child abuse has occurred.
The Allegheny Family Screening Tool — the AFST, one of the most studied predictive algorithms in American child welfare, running on maltreatment referrals in Allegheny County, Pennsylvania, since August 2016 — produces a Family Screening Score from 1 to 20 to assist hotline staff.
That score is not a measured probability that a child is being harmed.
At referral, the system calculates child-level risks associated with future re-referral and out-of-home placement, then produces a Family Screening Score that helps inform the hotline screening decision.
The distinction is not pedantic.
It is the whole game.
Child maltreatment is largely unobserved — no dataset records what actually happens inside every home. Administrative outcomes, by contrast, are recorded because the agencies themselves generate those records.
A model trained on historical administrative outcomes is therefore learning from what the system has previously observed and done. That creates a serious concern: patterns embedded in past investigations, interventions, and placements can be reproduced in later predictions, including patterns created by unequal surveillance or intervention.
The American Civil Liberties Union said this plainly in its 2021 report Family Surveillance by Algorithm: any tool built from historical jurisdictional data carries the risk of reproducing biases embedded in those data.
The pattern repeats wherever these systems appear.
Douglas County, Colorado’s tool — the Douglas County Decision Aid, built by the same broader research network associated with the AFST — predicts a child’s likelihood of future system involvement, including removal.
Allegheny’s separate Hello Baby model, applied to newborns, is designed to identify infants assessed as being at greatest risk of later child-welfare involvement.
In each case, the question that matters is what outcome the model actually predicts — not what shorthand is used to describe it publicly.
Once you see the substitution, the possibility of a feedback loop becomes clear.
Past agency decisions become training data. Model outputs can influence later agency decisions. Those later decisions may eventually become part of the next generation of administrative data.
A family that resembles families heavily represented in historical agency records may therefore attract heightened attention not because anyone directly observed danger, but because the system recognizes patterns already embedded in its own record.
Who sees the number — and who never does
Now trace the score’s path through a single case, because where the score appears — and where it does not — matters.
In Allegheny County, the Family Screening Score is generated at hotline screening and shown to the staff deciding whether a referral should be investigated.
Allegheny County states that the score is not shared beyond call screening.
That means the score can influence an upstream decision without itself becoming a downstream child-welfare determination.
The family may never see the number that helped inform the screening decision, and the score itself does not travel forward as evidence in the case.
So the machine’s influence enters at the front door — helping inform which referrals are investigated — and then does not continue forward as a separate child-welfare decision.
The decision is visible.
The internal screening input may not be.
Keep that architecture in mind through everything that follows, because it raises the central accountability question: these tools can influence consequential screening decisions without a uniform statutory framework requiring disclosure to affected families, independent predeployment review, or recurring bias audits.
Individual jurisdictions, including Allegheny County, have nevertheless commissioned and published evaluations.
It is difficult to contest a number you have never been told exists.
THE INPUTS
A POVERTY DETECTOR BY DESIGN
What the machine eats
If the target variable is the first question, the feature list is the second.
The AFST draws on extensive administrative records housed within Allegheny County’s integrated data system, including prior child-welfare involvement and information connected to other public systems. Versions of the tool have incorporated data associated with healthcare, behavioral health, public benefits, criminal-legal contact, and other government services.
Read that list again and notice what many of those data sources have in common: they disproportionately capture people whose lives intersect with public-benefit, public-health, disability, criminal-legal, or child-welfare systems.
A middle-class parent who sees a therapist through an employer health plan may leave a very different administrative footprint from a parent receiving behavioral-health treatment through Medicaid.
A parent with a private disability policy may appear differently in government records from a parent receiving SSI.
The model is not directly measuring parenting.
It is measuring patterns contained in administrative records, and the density of those records can correlate with poverty, disability, race, and prior contact with government systems.
The Human Rights Data Analysis Group and the ACLU took this from theory to measurement in a 2023 paper presented at the ACM Conference on Fairness, Accountability, and Transparency.
Working at the level of individual predictors associated with the AFST, the researchers found significant racial differences in exposure to some variables. One juvenile-probation-related predictor, for example, affected substantially more referrals involving Black household members than referrals involving non-Black households.
That does not prove the model is measuring race directly.
It does show how variables created by unequal contact with public systems can function inside a risk score.
New York City runs one of the least transparent versions examined here.
In May 2025, the investigative outlet The Markup reported that the city’s Administration for Children’s Services has, since 2018, used a model containing 279 variables — including neighborhood, mother’s age, number of siblings, and mental-health history — to identify already-open investigations for additional quality-assurance review by a higher-level specialist team.
Families are not told when the algorithm flags their case.
Neither are their lawyers.
Caseworkers themselves are not necessarily told that a case was selected algorithmically for the added review.
ACS’s own technical review acknowledged the presence of implicit and systemic bias in the underlying data while concluding that the model performed better than earlier criteria.
This, in a city where Black families experience dramatically greater contact with the child-welfare system than white families.
A 279-variable undisclosed score operating inside that system deserves scrutiny even when it is not itself the mechanism that opens the original investigation.
Why the inputs meet the mission
Now put the inputs beside the American definition of neglect, because the two facts matter together.
Neglect is the largest maltreatment category in federal child-welfare data. In federal fiscal year 2023, 546,159 children were identified as victims of maltreatment nationally, with neglect accounting for the largest category.
And “neglect,” in day-to-day American practice, can include conditions strongly affected by economic hardship: an empty refrigerator, unstable housing, utility loss, or unmet medical needs.
But neglect is broader than poverty, and the two should not be treated as synonymous.
Feed that system a model whose inputs can include public-benefits history, public healthcare, behavioral-health records, and prior government-system contact, and a serious problem emerges: a tool intended to identify child-welfare risk can also reproduce patterns of poverty, disability, and prior surveillance.
The next question is unavoidable:
What happened in the places that actually switched one on?
That is Part II.
YOUR MOVE:
FIND OUT WHAT YOUR STATE USES
You do not need to know who to call yet.
Start here.
Open Google and search:
[your state] child welfare agency
Look for the official
.govresult. Depending on your state, the agency may be called Child Protective Services, Department of Children and Families, Department of Human Services, or something similar.Once you are on the official agency website, use its search box and try these terms one at a time:
predictive analytics
risk assessment
risk scoring
screening tool
decision support
algorithm
automated decision-making
Then Google the agency name with the same terms:
“[agency name]” predictive analytics
“[agency name]” risk scoring
“[agency name]” algorithm child welfareSometimes contracts, government reports, meeting minutes, audits, or news coverage are easier to find through a regular web search than through the agency’s own website.
If you find the name of a tool, write it down.
Then search:
“[tool name]” child welfare
Look for:
what the tool predicts
what information it uses
who built it
whether families are told it is being used
whether an independent evaluation has been published
If you cannot find an answer, that is when you locate the agency’s Public Information, Open Records, or FOIA/Public Records page.
You do not need to understand public-records law yet. For now, simply find the page and save it.
Your first assignment is not to become an expert.
It is to answer one question:
Does my state or county use a score?
If you find one, save the name, the link, and a screenshot.
Because silence protects systems.
Informed communities protect children.
If you enjoyed this work and feel encouraged by independent investigations like this, please consider becoming a paid subscriber to https://17sog.substack.com here, or donate at Project Milk Carton Inc. or at https://t.co/PvtnrODkNz
Your support helps us turn information into action for children and families by:
keeping high-risk and missing children visible
identifying gaps in the systems responsible for protecting them
giving families and communities clear, usable information they can act on
supporting education, outreach, and practical community-based solutions
maintaining the public-data and OSINT work that makes this accountability possible
building tools and resources designed to help communities protect children before they disappear into a system
Most importantly, please share this article.
Transparency survives when ordinary people help spread information.
And informed communities protect children.
ALL paid subscriptions are 100% tax deductible
Project Milk Carton | 501(c)(3) | EIN: 33-1323547
A NOTE ON SOURCES — AND WHAT WE COULD NOT CONFIRM
This installment is built on Allegheny County documentation concerning the Allegheny Family Screening Tool and related predictive-risk systems; public documentation concerning Hello Baby and the Douglas County Decision Aid; the ACLU’s Family Surveillance by Algorithm (2021); ACLU and Human Rights Data Analysis Group research presented at ACM FAccT 2023; The Markup’s May 2025 investigation of New York City’s ACS model; federal child-maltreatment data; and the federal funding opportunity HHS-2026-ACF-ACYF-CA-0037, posted May 28, 2026 and modified June 12, 2026.
Where the record requires qualification, we have done so:
The Allegheny Family Screening Tool does not simply predict whether a child will be placed in foster care. It incorporates predicted risks associated with future re-referral and out-of-home placement into the Family Screening Score used during hotline screening.
New York City’s ACS model does not determine which families receive an initial investigation. It is used to identify already-open investigations for additional quality-assurance review.
Allegheny County states that its Family Screening Score is not shared beyond call screening. We therefore distinguish between a score influencing an upstream screening decision and a score becoming evidence later in a dependency case.
An Indiana statistic included in an earlier draft stated that 87 percent of children entering foster care in 2024 had parents who were not accused of physical or sexual abuse. We could not tie that figure to the underlying state report during this review, so it has been omitted pending verification.
Civil-rights complaints alleging disability discrimination involving the AFST were filed in 2022. Associated Press reporting in January 2023 said Justice Department civil-rights attorneys were scrutinizing the tool. We found no publicly announced findings, settlement, enforcement action, or disposition and therefore describe the public record as unresolved rather than characterizing the matter as a confirmed open federal investigation.
Project Milk Carton is a 501(c)(3) transparency and public-data organization. This article is public-record research, not legal advice.
Corrections: if any figure or citation here is wrong, we will say so, prominently.
Next: The Number That Knocks — Part II: The Record
Because silence protects systems.
Informed communities protect children.
• projectmilkcarton.org • t.me/ProjectMilkCarton